Gender Inequality in New Media: Evidence from

Transcription

Gender Inequality in New Media: Evidence from
Wikipedia∗
Marit Hinnosaar†
May 2015
Abstract
Media is considered to be critical for gender equality. I analyze Wikipedia, one of
the prominent examples of new media. I study why women are less likely to contribute
to Wikipedia, the implications of the gender gap, and what can be done about it. I
find that: (1) gender differences in the frequency of Wikipedia use and in beliefs about
one’s competence explain a large share of the gender gap in Wikipedia writing; (2)
the gender gap among contributors leads to unequal coverage of topics; (3) providing
information about gender inequality has a large effect on contributions.
JEL codes: L86; L82; J16; H41
Keywords: Gender, Internet, Media, Public goods
1
Introduction
United Nations is engaged in an initiative to increase female presence in the media.1 The
initiative was established “in recognition that the media are critical to the achievement of
gender equality and women’s empowerment”.2 In most traditional media outlets, men vastly
outnumber women, and several major media organizations have taken steps to change that.
Examples include Bloomberg introducing gender quotas for news stories3 and BBC for panel
∗
I’m grateful to Ainhoa Aparicio, Gabriele Camera, Matthias Doepke, Martin Dufwenberg, Avi Goldfarb,
Shane Greenstein, Toomas Hinnosaar, Ryan McDevitt, David McKenzie, Mario Pagliero, Andrea Prat,
Stephan Seiler, Alex Tetenov, and Michael Zhang for helpful comments and discussions.
†
Collegio Carlo Alberto, [email protected].
1
http://www.unesco.org/new/en/communication-and-information/crosscutting-priorities/
gender-and-media/global-alliance-on-media-and-gender/about-gamag/, accessed May 15, 2015.
2
The Global Alliance on Media and Gender. Framework and Plan of Action. http://www.unesco.
org/new/fileadmin/MULTIMEDIA/HQ/CI/CI/pdf/gamag_framework_and_action_plan_30sep2014.pdf,
accessed May 15, 2015.
3
Posetti, Julie. 2015. “The Bloomberg News recipe for newsroom transformation,” World News Publishing
Focus by World Association of Newspapers and News Publishers, March 17. http://blog.wan-ifra.org/
2015/03/17/the-bloomberg-news-recipe-for-newsroom-transformation.
1
shows.4 One reason to introduce these measures, is that historically rigid institutions, work
arrangements, and hiring policies in media, might otherwise inhibit women to be heard. In
this paper, I analyze a new media source, which is not restricted by these institutions, but
which nevertheless faces similar issues.
I analyze Wikipedia, which is a widely used encyclopedia and one of the most prominent
examples of collective knowledge production and new media. It is the sixth most popular
website both in the US and in the world5 and many people rely on its information.6 Although
anyone can write and edit Wikipedia entries, there is evidence that the majority of Wikipedia
writers are men. According to surveys conducted among Wikipedia contributors, only
9 percent7 to 16 percent8 of them are women. The Wikimedia Foundation, which owns
Wikipedia, views the lack of female contributors as a problem and has set itself goals to
increase their share.9
Why does it matter who contributes to Wikipedia? Who contributes to its production
matters if a gender gap among contributors leads to biased content. There is some evidence
that Wikipedia’s coverage of topics is gender biased. In particular, popular media has voiced
concerns that Wikipedia has less information about female than male public figures.10 For
example, consider the authors whose fiction books reached number one on New York Times’s
”Best Sellers” list for the first time between 2005 and 2014. Wikipedia articles about such
male authors are on average 88 percent longer than those of female authors.11 More generally,
Reagle and Rhue (2011) document that biographies of women are more likely to be missing
4
BBC. 2014. “BBC TV boss: No all-male panel shows”, February 9. http://www.bbc.com/news/
entertainment-arts-26107011.
5
The popularity of Wikipedia is measured by the web traffic measurement company Alexa Internet
(http://www.alexa.com/siteinfo/wikipedia.org, accessed March 16, 2015).
6
Many people unknowingly access information provided by Wikipedia because online services, such as
Google and Apple’s Siri, use it. Google’s Knowledge Graph display (the box on the right-hand side of Google’s
search results) is often based on Wikipedia (http://en.wikipedia.org/wiki/Knowledge_Graph, accessed
January 9, 2015). http://www.siriuserguide.com/siri-faq/, accessed January 9, 2015.
7
Wikimedia Foundation. 2011. ”Editor Survey,” April. http://upload.wikimedia.org/wikipedia/
commons/7/76/Editor_Survey_Report_-_April_2011.pdf; Wikimedia Foundation. 2011. ”Second
2011 Wikipedia Editor Survey,” December. https://upload.wikimedia.org/wikipedia/commons/8/84/
December_2011_Wikipedia_Editor_Survey_topline.pdf.
8
Glott, Ruediger; Schmidt, Philipp; Ghosh, Rishab. 2010. ”Wikipedia Survey: Overview
Results”, March. http://web.archive.org/web/20100414165445/http://wikipediasurvey.org/docs/
Wikipedia_Overview_15March2010-FINAL.pdf.
9
BBC. 2014. ”Wikipedia ’completely failed’ to fix gender imbalance,” BBC interview with Jimmy Wales,
August 8. http://www.bbc.com/news/business-28701772.
10
Cohen, Noam. 2011. ”Define Gender Gap? Look Up Wikipedia’s Contributor List,” New York Times,
January 30. http://www.nytimes.com/2011/01/31/business/media/31link.html.
11
From 2005 to 2014, there were 32 female authors and 28 male authors whose fiction books reached number
one on New York Times’s ”Best Sellers” list and whose previous books had not been at the position. Of
these bestselling books by female authors, 47 percent don’t have a Wikipedia page, but only 19 percent of
the books by male authors don’t have a page.
2
from Wikipedia compared to Encyclopedia Britannica than those of men. Coverage of public
figures is important because they are role models. Literature provides evidence that role
models influence educational attainment (Beaman, Duflo, Pande, and Topalova, 2012; Fairlie,
Hoffmann, and Oreopoulos, 2014) and political aspirations (Wolbrecht and Campbell, 2007)
and that exposure to female leaders weakens stereotypes about gender roles and reduces
gender bias (Beaman, Chattopadhyay, Duflo, Pande, and Topalova, 2009).
In this paper, I study gender inequality in new media. My goal is to learn why women
are less represented and how to change that. I restrict attention to Wikipedia, a prominent
example of new media. Specifically, the questions I ask are the following: (1) Why are women
less likely to contribute to Wikipedia? (2) What are the implications of this gender gap on
Wikipedia’s content? (3) What can be done about it? The analysis also sheds some light to
gender inequality in traditional media. It highlights the importance of representing diverse
opinions and informs about measures to achieve that.
To answer these questions, I use a large dataset of Wikipedia editing histories, and I
conduct a survey and a randomized survey experiment with 1,000 participants on Amazon
Mechanical Turk. The dataset of Wikipedia editing histories includes all registered editors
who have reported their gender. I analyze their editing behavior on almost 900,000 biographies
in the English-language edition of Wikipedia. I use the data to learn whether female editors
are more likely than male editors to edit biographies of women. To complement the dataset,
I conducted an online survey and a randomized survey experiment with 1,000 people on
Amazon Mechanical Turk.12 This sample consists of internet users who could potentially
contribute to Wikipedia (and some of them have done that), which allows me to study why
individuals do or do not contribute to Wikipedia. The goal of the survey was to understand
which factors influence Wikipedia editing behavior and how Wikipedia editing behavior is
related to contributions to other online and offline public goods. The goal of the experiment
was to test whether providing information about gender inequality in Wikipedia changes
editing behavior, specifically, whether the information provided to the treatment group (1)
increases the likelihood of editing Wikipedia articles about women and (2) increases the
likelihood that survey participants, especially women, edit Wikipedia in the future.
It has been puzzling that women contribute less to Wikipedia than men. There are no
gender differences in the amount of free time.13 The gender gap in Internet usage is small,
and there are no large gender differences in online behaviors, such as social networking,
online news consumption, communication, and e-commerce.14 Moreover, women contribute
12
The randomized survey experiment is registered at the AEA RCT Registry with the number AEARCTR0000500.
13
Aguiar and Hurst (2007) document that in the US, men and women allocate about equal time to leisure.
14
Of US adults, 87 percent of men and 86 of women used the Internet in 2014 according to the Pew
3
no less than men to another example of online public good provision, writing user reviews
for products and services.15 Based on the survey, I find that almost half of the gender gap
in Wikipedia writing is explained by gender differences in two characteristics: frequency of
Wikipedia use and belief about one’s competence.
The fact that most contributors are men could lead to biased coverage in Wikipedia.
I look at the articles about human beings. Based on both the survey and the dataset of
Wikipedia editing histories, I find that women are about twice as likely as men to contribute
to Wikipedia articles about women. The result continues to hold when restricting attention
to specific groups of people, such as professionals or cultural figures. I analyze whether the
distribution of editorial input reflects demand for the articles among readers. However, I find
no evidence that articles about women have lower demand than those about men. On the
contrary, the number of readers per editor is higher for articles about women, and the share
of articles that no one reads is larger in the case of articles about men.
What can be done about gender inequality in Wikipedia? From a randomized survey
experiment, I find that providing information about gender inequality in Wikipedia almost
doubles the likelihood of contributing to articles about women. However, the treatment
of providing information decreases men’s likelihood of contributing to Wikipedia and has
no effect on women. That is, the treatment leads editors to redirect their editorial input
and equalizes participation among men and women but decreases overall participation in
Wikipedia editing. Because there are other ways to increase overall contributions to Wikipedia,
the informational treatment can still provide a useful way to decrease gender inequality.
In conducting the randomized survey experiment, I follow the approach taken by Kuziemko,
Norton, Saez, and Stantcheva (2015), who studied how information about inequality changes
reported preferences for redistribution in a similar research design on Amazon Mechanical
Turk. Amazon Mechanical Turk has recently gained popularity as a platform for experiments
in economics (for other examples, see Horton, Rand, and Zeckhauser, 2011). Methodological
papers have evaluated the platform’s suitability for surveys and experiments (see, for example,
Research Center (”The Web at 25.” 2014, Pew Research Center, February 28. http://www.pewinternet.
org/2014/02/25/the-web-at-25-in-the-u-s). In the UK in 2014, 82 percent of men and 81 percent of
women between 16 and 74 years old used the Internet daily according to Eurostat. In the same sample in the
UK, more women participated in online social networks (62 percent of women versus 58 of men), and more
men read online news (64 percent of men versus 55 of women). In other activities, such as sending emails,
making online purchases, and doing online banking, the differences are even smaller. (”Information Society
Statistics,” Eurostat. http://ec.europa.eu/eurostat/web/information-society/data/database.)
15
Chen, Harper, Konstan, and Li (2010) found that women contribute more to online movie ratings. Surveys
by Pew Research Center find no gender difference in writing user reviews (”Health Online 2013,” 2013, Pew
Research Center. http://www.pewinternet.org/files/old-media//Files/Reports/PIP_HealthOnline.
pdf; ”Online Product Research,” 2010, Pew Research Center. http://www.pewinternet.org/files/
old-media//Files/Reports/2010/PIP%20Online%20Product%20Research%20final.pdf).
4
Mason and Suri, 2012), and I followed their advice.
The paper contributes to an emerging field in economics that studies contributions to new
media. Contributions to Wikipedia have been studied by Zhang and Zhu (2011), Aaltonen
and Seiler (2015), and Algan, Benkler, Morell, and Hergueux (2013).16,17 Zhang and Zhu
(2011) studied the relationship between contributions to Wikipedia and group size using an
exogeneous decrease in group size in Chinese-language Wikipedia. Aaltonen and Seiler (2015)
studied the impact of the cumulative growth of content on new contributions. Algan, Benkler,
Morell, and Hergueux (2013) conducted online experiments with current Wikipedia editors
to test whether pro-social motives explain their contributions to Wikipedia. None of these
papers examined the role of gender in contributions, which is the subject of this paper.
More generally, the paper contributes to empirical studies of media bias. These have
concentrated mainly on measuring media bias (Groseclose and Milyo, 2005) and analyzing
sources of bias, including how consumer demand affects news (Gentzkow and Shapiro, 2010),
how online, user-generated content affects print newspapers (Yildirim, Gal-Or, and Geylani,
2013), and advertisers’ influence on news content in politics (Tella and Franceschelli, 2011),
in financial media (Reuter and Zitzewitz, 2006), and in blogs (Sun and Zhu, 2013). This
paper studies bias in a relatively new media source, Wikipedia. Bias in Wikipedia has been
studied in the case of Democrat versus Republican bias in political articles (Greenstein and
Zhu, 2012, 2014) and gender bias in biographies (Reagle and Rhue, 2011). Reagle and Rhue
(2011) compared biographies in Wikipedia and Encyclopedia Britannica and documented
that Wikipedia biographies of women are more likely to be missing than biographies of men
relative to Britannica. Although they measured the bias in coverage, they did not analyze
the sources of bias or how to overcome it, which are the main questions studied in my paper.
This paper also relates to literature on gender economics, in particular gender differences
in contributions in a group, confidence, and the provision of public goods. The results
presented in the paper are consistent with the findings of Coffman (2014), who shows that
women are less likely to contribute their ideas to a group. I find that one of the reasons
women contribute less to Wikipedia is differences in beliefs about one’s own competence.18
This is related to the results on gender and competition (for an overview, see Niederle and
Vesterlund, 2011), where the difference in confidence is found to be one of the reasons for
gender difference in competitiveness. Interestingly, literature on public good provision does
16
Contributions to other forms of new media have been studied, for example, in the case of user reviews
(e.g., Chen, Harper, Konstan, and Li, 2010), Twitter (e.g., Toubia and Stephen, 2013), mobile-phone-based
content generation (Ghose and Han, 2011), and local information provision to an extreme sports website
(Shriver, Nair, and Hofstetter, 2013).
17
Data from Wikipedia has also been used to study questions that are less related to this paper, such as
the impact of copyright (Nagaraj, 2014).
18
However, in this paper, we cannot confirm whether the beliefs about one’s competence are correct.
5
not provide a clear reason for why we should expect women to contribute less to Wikipedia.
Experimental evidence on the role of gender in public good provision and, more generally,
social preferences is mixed (for overviews, see Eckel and Grossman, 2008; Croson and Gneezy,
2009; Bertrand, 2011).
This paper does not analyze how gender inequality in new media affects real-life decisions.
However, the above cited literature on the importance of role models suggests that it might
have an impact on educational choices and political aspirations. Additionally, several studies
in media economics provide evidence that media matters. For example, local television news
affected voter turnout in the US (Oberholzer-Gee and Waldfogel, 2009), voting decisions were
affected by candidate endorsements by newspapers in the US (Chiang and Knight, 2011)
and by independent TV in Russia (Enikolopov, Petrova, and Zhuravskaya, 2011), access
to cross-border radio had an impact on nationalism in Croatia (DellaVigna, Enikolopov,
Mironova, Petrova, and Zhuravskaya, 2014), cable television affected women’s status in India
(Jensen and Oster, 2009), and soap operas affected fertility decisions in Brazil (Ferrara, Chong,
and Duryea, 2012).
The remainder of the paper is organized as follows. Section 2 describes the Wikipedia’s
editing process, the dataset of Wikipedia editing histories, the survey, and the randomized
survey experiment. Section 3 presents empirical results that answer three questions. First,
why are women less likely to edit Wikipedia compared to men? Second, how does the gender
gap in participation affect biographies on Wikipedia? Third, how does the treatment of
providing information affect gender inequality in Wikipedia? Section 4 concludes.
2
Background and data
2.1
Wikipedia’s editing process
Wikipedia is a free-access Internet encyclopedia. As of January 1, 2015, it contains 34 million
articles, including 4.7 million articles in English-language Wikipedia.19 Wikipedia is owned
by the Wikimedia Foundation, a non-profit foundation. It is written by volunteers.
Anyone can create Wikipedia articles and edit almost any of its existing articles. There
are two types of editors: anonymous and registered. Registration is not required for most
editing activities. However, to create new articles or to upload images, editors have to register.
For each article, Wikipedia records who, when, and how edited the article. In the case of
anonymous editors, their computers’ IP addresses are recorded.
19
http://en.wikipedia.org/wiki/Wikipedia, accessed January 1, 2015.
6
Wikipedia’s editing community is large. According to Wikimedia Statistics,20 as of
December 2014, about 5.3 million registered editors had contributed to the English-language
edition of Wikipedia. Most editors made only a few edits, but about 976,000 editors had
edited at least 10 times. Registered editors made about 61 percent of edits in English-language
Wikipedia.
2.2
Dataset of Wikipedia editing histories
2.2.1
Dataset construction
The dataset is constructed from editing histories of the English-language edition of Wikipedia.
To analyze editing by the gender of the editor, I restrict attention to the editors who have
reported their gender when registering their Wikipedia username. I analyze their editing
behavior of articles about human beings. In particular, I’m interested in whether they edit
articles about men or women.
To learn whether a Wikipedia article is about a human being, I use information from
Wikidata, which is a database managed by the Wikimedia Foundation. From Wikidata, I also
learn the gender and year of birth of the human being. I include in the sample all the human
beings in Wikidata as of July 2014 who were born in any year from 1000 to 2000 and who
were either male or female. There are almost one million such human beings. When there
was more than one birth year in the database, the earliest was taken (there were 1,445 such
human beings in the database). I exclude all human beings whose gender is not determined
to be either male or female in Wikidata (either gender is missing, it is not male or female,
or the person has more than one gender in the database), there are 41 such cases. After
excluding those, we are left with the Wikidata final sample of 924,371 human beings.
From the Wikidata database, I also obtain information about the professions of the human
beings. I use the information to look at subgroups and exclude celebrities like actors, athletes,
and pop stars. Of the sample of human beings selected above, 490,023 have at least one
profession in the Wikidata database. I categorize all the professions that are held by at least
1,000 human beings in the sample, which is 88% of all the professions held by these people,
into three groups. The first group I call Professionals, this includes lawyers, politicians, and
scientists among others. The second group I call Culture, where I include writers, painters,
composers, etc. but exclude actors and singers. The third group called Other consists mostly
of athletes but also includes actors and singers. The exact list is in Appendix C in Table C.1.
As of July 2014, 259,638 registered editors of English-language Wikipedia had reported
their gender when registering their Wikipedia username. Of those, 14.4 percent reported their
20
http://stats.wikimedia.org/EN/TablesWikipediaEN.htm, accessed January 31, 2015.
7
gender as female. The number of editors who reported their gender is not small; however, the
majority of registered editors have not reported their gender. This is not surprising because
Wikipedia requires almost no information to register, no real name or even an email address.
That only a small percentage of editors reports their gender is one of the reasons why I
complement the Wikipedia editing histories with the survey data.
The editors who report their gender tend to be more active, with a larger number of edits.
Note that for my analysis it is not important that male and female editors report their gender
at equal rates. For example, it might be that women are less likely to report their gender
for fear of harassment. For this analysis that is not a problem. What is important is that
editors don’t systematically report wrong gender. Since the default option is not specifying
one’s gender, I would not expect that they are massively reporting wrong gender.
The sample of editors used in the paper consists of all the editors who reported their
gender and who have edited English-language Wikipedia articles about human beings in the
Wikidata final sample. There are 74,817 such editors. They have edited 866,431 articles
about the human beings in the Wikidata final sample.
2.2.2
Summary statistics
Table 1 presents summary statistics of the Wikipedia editors in the sample. 12.6 percent of
editors are women. Women contribute somewhat longer texts than men, both per article
and in total per year. Changes over time in the number of active editors and the share of
female editors are depicted in Figure 1. The percentage of female editors has increased from
3.7 percent in 2002 to a peak of 11.5 percent in 2011. In 2013, 10.4 percent of the active
editors were female. Changes in the percentage of female editors mirrored changes in the
total number of active editors. The number of active editors increased monotonically from 81
in 2002 to 27,600 in 2011 and was 22,600 in 2013.
2.3
Survey and experiment
The survey together with the randomized survey experiment was administered in September
2014. The survey had the following structure: (1) background socioeconomic questions
including gender, age, education, work status, and children; (2) questions about previous
Wikipedia use and editing; (3) randomized treatment providing information about gender
inequality in Wikipedia shown only to the treatment group; the control group was shown
information about Wikipedia that was not related to gender; (4) hypothetical situation where
one is asked to edit Wikipedia; specifically, respondents were asked to choose a Wikipedia
page of a person and find some information from the Internet that is missing from the
8
Wikipedia page; note that the respondents were not asked to actually edit Wikipedia; (5)
questions about Wikipedia editing, including respondents’ beliefs about their competence
and the likelihood of editing Wikipedia in the future; and (6) questions about charitable
giving, volunteering, writing user reviews, and free time.
2.3.1
Randomized survey experiment
The goal of the experiment was to test whether providing information about gender inequality
in Wikipedia changes editing behavior. Individuals were randomly assigned to either a treatment or a control group with equal probabilities. The information provided to the treatment
group is a quote from the Wikipedia page, ”Gender Bias on Wikipedia”21 : ”Wikipedia
has been criticized by some academics and journalists for having only 9% to 13% female
contributors and for having fewer and less extensive articles about women or topics important
to women.” The control group read the following information about Wikipedia, which has
nothing to do with gender bias: ”Wikipedia started in 2001. English-language Wikipedia has
over 4.5 million articles.”
After reading the information, respondents were asked to imagine a hypothetical situation
in which they edit a person’s Wikipedia page. Respondents were asked to look at Wikipedia
articles and find some relevant information from the web that is missing from a Wikipedia
article. Note that respondents were never asked to actually edit Wikipedia. Respondents
were only asked to report the Wikipedia page they would choose and the information they
would hypothetically add to the page. In the end, they were also asked how likely they are
to edit Wikipedia in the future.
2.3.2
Data collection
The survey was posted on Amazon Mechanical Turk. Amazon Mechanical Turk is an Internet
marketplace, where workers perform tasks, including answering surveys and participating in
experiments. The survey was described as a 20-minute research study on Wikipedia, which
pays the participant $1.50. The payment gives an effective hourly wage close to the average
effective hourly wage on Amazon Mechanical Turk according to Kuziemko, Norton, Saez, and
Stantcheva (2015).
To ensure high-quality responses, several measures were taken based on advice from
previous studies that used Amazon Mechanical Turk. First, only workers with good a track
record were allowed to take the survey. The exact qualification criteria were that a worker
21
The Wikipedia page, ”Gender Bias on Wikipedia”, was from August 24, 2014: http://en.wikipedia.
org/w/index.php?title=Gender_bias_on_Wikipedia&oldid=622670577
9
must have completed and approved at least 50 tasks, and his or her approval rate must be
95 percent or higher. Second, only workers from the US were eligible to take the survey. I
chose to limit the set of possible respondents to those in somewhat similar environments
because the survey asked about volunteering and writing user reviews. The location was
checked and restricted by Amazon itself. According to computer IP addresses of the workers
who completed the survey, a very small percentage came from locations outside of the US.
Third, respondents were told that payment would be contingent on completing the survey
and providing a survey code visible only after finishing. The respondents were not allowed to
skip any questions. Fourth, respondents were told that they can start the survey only once.
The main concern was that they reach the treatment page multiple times and see both the
treatment group’s and control group’s information. One respondent reached the treatment
page twice, but by chance, he received the same treatment both times, so I still include his
responses in the sample.
The survey stayed open until 1,000 workers reported that they had completed it. Of the
1,000 Amazon Mechanical Turk workers, 26 did not provide a valid survey response: four
workers could not be matched to any survey response (because they either had not started
the survey or did not provide enough information to match to the survey response), three
started but did not complete the survey, and 19 did not provide a link to the Wikipedia page
of a person (because they either misunderstood the question or ignored the instructions).
After excluding these respondents, 974 remained.
2.3.3
Summary statistics
Table 2 shows characteristics of the 974 respondents who completed the survey. In terms
of demographic characteristics, the sample is quite representative of the general population:
49.5 percent of the sample is female, 54.0 percent has at least a college degree, 57.7 percent
works at least 35 hours per week, and 10.0 percent is unemployed. However, the sample is on
average younger, only 35.4 percent is at least 35 years old. This is similar to other studies
using Amazon Mechanical Turk, including Kuziemko, Norton, Saez, and Stantcheva (2015).
Compared to a representative sample of U.S. adults, the population on Amazon Mechanical
Turk is restricted to Internet users. However, this fits well the purpose of my study, because
these individuals edit or potentially could edit Wikipedia.
21.9 percent of the sample has edited Wikipedia previously. Of those who have edited,
31.5 percent are women. It is difficult to say how close is this to a representative sample of
U.S. Internet users, because anyone can edit Wikipedia without registering. A back-of-theenvelope calculation says that in 2008, 16 percent of Internet users had edited Wikipedia.
This calculation is subject to caveats, because it combines information from two different
10
surveys in 2008, according to which: 47 percent of U.S. Internet users used Wikipedia22 and
33 percent of Wikipedia users had edited Wikipedia.23
Table 3 presents respondents characteristics for men and women in the control group
separately by gender. The treatment group is excluded here to make sure the responses are
not affected by the treatment. In terms of demographic characteristics, there are significant
differences between men and women only in employment: 65 percent of men and 46 percent
of women work at least 35 hours per week. There are several significant differences between
men and women in terms of Wikipedia-related behavior and attitudes. Women use and edit
Wikipedia significantly less than men. Note that the frequency of Wikipedia use depends on
both the frequency of using encyclopedias and preference for Wikipedia over other sources,
such as Encyclopedia Britannica. A larger share of women believe they are less competent
than other editors to edit Wikipedia. There are no significant differences between men and
women in writing user reviews, in the amount of free time, and in volunteering. However,
women are significantly more likely than men to give to charity.
2.4
Additional data sources
I use data on Wikipedia viewership statistics to describe pages’ popularity. The dataset
includes for each Wikipedia page the number of page views in September 2014.24 Page views
are from a time period outside the editing histories in the sample. On average, the number of
page views is rather stable across months, and there is no reason to believe that September
2014 is somehow different from other time periods.
3
Empirical results
This section provides empirical results that answer three questions. First, why are women
less likely to contribute to Wikipedia? Second, what are the implications of the gender gap
in participation on the information available on Wikipedia? Third, how does the treatment
of providing information about gender inequality in Wikipedia affect editing behavior?
22
Pew Internet & American Life Project Poll, Nov, 2008. http://www.ropercenter.uconn.edu/psearch/
question_view.cfm?qid=1727230&pid=53&ccid=50, accessed May 14, 2015.
23
Glott, Ruediger; Schmidt, Philipp; Ghosh, Rishab. 2010. ”Wikipedia Survey: Overview
Results”, March. http://web.archive.org/web/20100414165445/http://wikipediasurvey.org/docs/
Wikipedia_Overview_15March2010-FINAL.pdf.
24
The data on page views is downloaded from: http://dumps.wikimedia.org/other/pagecounts-raw/
11
3.1
Why are women less likely to edit Wikipedia?
To answer the question why women contribute less to Wikipedia, I use the survey data. I
present two sets of results. First, I analyze which characteristics and behaviors are correlated
with editing Wikipedia. Second, I analyze how much of the gender difference in Wikipedia
editing is explained by the gender differences in these characteristics. I use two measures
for Wikipedia editing: (1) whether the respondent has edited Wikipedia in the past and
(2) whether the respondent reports that he or she is likely to edit Wikipedia in the future.
Throughout the analysis, the sample is restricted to the control group in the survey experiment.
Table 4 presents results from logit regressions where the dependent variable, an indicator
for editing Wikipedia, is regressed on gender and other characteristics. These include
demographic characteristics (indicators for being 35 or older, having a college degree, being
unemployed, and working at least 35 hours per week), behaviors and attitudes toward
Wikipedia (indicators for using Wikipedia daily, using Wikipedia at least weekly, and belief
about one’s competence compared to other Wikipedia editors), and other behaviors (indicators
for donating to charity in the past year, volunteering at least one hour per week in the past
year, writing a user review in the past six months, and having at least three hours of leisure
time per weekday).
In column 1, the dependent variable is an indicator for whether the respondent has edited
Wikipedia in the past. Estimates show that those who use Wikipedia daily or who have
written a user review in the past six months are more likely to have edited Wikipedia. A
stronger belief about one’s competence compared to other potential editors is positively
associated with the likelihood of having edited Wikipedia. The gender difference in the
belief about competence could be due to women being less competent or due to women
underestimating their competence. The survey data does not allow to distinguish these.
In column 2, the dependent variable is an indicator for whether the respondent reports he
or she is likely to edit Wikipedia in the next 30 days. Specifically, the variable takes value
one if the respondent answered that he or she is either “very likely” or “quite likely” to the
question: “How likely or unlikely is it that you will actually edit Wikipedia over the next
30 days?” The only characteristics that are statistically significantly associated with the
likelihood of editing are daily use of Wikipedia and belief about one’s competence. Appendix
C presents Table C.2 with the same regressions estimated separately for men and women.
Let’s address the role of editor’s gender in the decision to edit Wikipedia. The estimates
in the first column of Table 4 imply that women are significantly less likely to have edited
Wikipedia in the past even controlling for other factors. But according to the estimates in
the second column, gender itself is not statistically significant, implying that the gender gap
in the likelihood of editing in the future is explained by other characteristics. To learn how
12
much of the gender difference in Wikipedia editing is explained by gender itself and how
much comes from men and women having different attitudes and behaviors toward Wikipedia,
I decompose the gender difference in editing Wikipedia using Oaxaca (1973) decomposition,
which was extended to non-linear equations by Fairlie (2005).25 For the explanatory variables,
I include indicators for daily Wikipedia use, at least weekly Wikipedia use, and belief about
one’s competence compared to other editors. The coefficient estimates of these variables
were statistically significant in the logit regressions in Table 4, and the variables have a large
gender difference as presented in Table 3.
Table 5 presents the results from the non-linear decomposition. In the first row, the
dependent variable is an indicator of whether the respondent has edited Wikipedia in the
past. The difference in editing rates of men and women is 15.0 percentage points. Gender
differences in the frequency of Wikipedia use and the belief about one’s competence explain
a large share, namely, 43 percent of the gender gap in editing. In the second row, the
dependent variable is whether the individual is likely to edit Wikipedia in the next 30 days.
The difference in editing rates of men and women is 5.2 percentage points. Gender differences
in these characteristics explain 79 percent of the gender gap in editing.
3.2
The effects of gender gap on Wikipedia content
In this subsection, I study whether gender gap among editors leads to gender inequality in the
biographical information available on Wikipedia. Specifically, I test whether female editors
are more likely than males to edit biographies of women. I use two alternative data sources:
(1) Wikipedia editing histories of editors who have reported their gender and (2) the survey
data.
One reason why more editorial input could be directed to the biographies of men is
demand for the articles among Wikipedia users. In the last part of the subsection, I study
25
ˆ can be written as
The decomposition for the non-linear equation Y = F (X β)
 M

W
N
X F (X M βˆP ) N
X F (X W βˆP )
i
i

Y¯ M − Y¯ W = 
−
M
W
N
N
i=1
i=1
 M
  W

M
W
N
N
X F (X M βˆM ) N
X F (X M βˆP )
X F (X W βˆW ) N
X F (X W βˆP )
i
i
i
i
+

+
−
−
M
M
W
W
N
N
N
N
i=1
i=1
i=1
i=1
where Y¯ j is the average probability of the binary outcome variable Y by gender j = M, W , N j is the sample
size by gender, F is the cumulative distribution function of the logistic distribution, Xij is a row vector of
explanatory variables of individual i of gender j, βˆj is a vector of coefficient estimates for gender j, and βˆP is
a vector of coefficient estimates from a pooled sample of both genders. The term in the first brackets captures
the part of the gender gap in editing that is explained by the gender differences in explanatory variables.
13
whether the allocation of editorial input is proportional to readership. Specifically, I analyze
whether there is less demand for existing articles about women compared to those about men.
Note that there could also be gender differences in demand for Wikipedia pages that do not
exist, but I concentrate only on demand for existing articles using data about Wikipedia
page views.
3.2.1
Evidence from Wikipedia editing histories
In this part, I simply document that Wikipedia female editors direct relatively more of
their editorial input to biographies of women. Table 6 presents the average (over editors)
percentage of contributions to biographies of women from contributions to all biographies.
An observation here is an editor. In column 2, the average percentage is calculated over male
editors and, in column 3, over female editors. Contributions can be measured in multiple
ways. In rows 1-3, contributions are measured in the length of text added.
The first row of Table 6 shows that the average percentage of text added to biographies
of women is 31% in the case of female editors and 17% in the case of male editors. To
make sure that the gender difference does not come only from celebrities like actors, athletes,
and pop stars, rows 2 and 3 restrict attention to subgroups of biographies. In row 2, the
sample includes lawyers, politicians, scientists, etc. (group Professionals); in row 3, it includes
writers, painters, composers, etc. (group Culture); the exact list of the groups is presented in
Appendix C, Table C.1. A large gender difference remains in both groups. Rows 4-6 measure
contributions by the number of biographies edited. The measure gives similar results.
Figure 2 presents the percentage of text added to biographies of women from text added to
all biographies, separately by male and female editors. Excluding the beginning of Wikipedia
when the number of editors was very small, the relative share of contributions to biographies
of women has increased, but the difference between male and female editors remained large.
In Appendix D, Figure D.1 presents the same statistics for the same subgroups of biographies
as above. The trends are similar to aggregate statistics.
3.2.2
Survey evidence
The dataset of Wikipedia editing histories is limited by the fact that editors must report their
gender, and relatively few editors do so. Therefore, I complement the analysis with survey
results. The survey data has two advantages. First, it represents a more general population.
Second, it includes more information about (potential) editors than only their gender. Hence,
it allows to test whether gender matters for the choice which article to edit, even when we
control for other characteristics of the editor.
14
This part presents two sets of results using survey data. First, I redo a similar exercise of
comparing the likelihood of editing biographies of women versus men by male and female
editors. The results largely confirm the ones from Wikipedia editing histories. Second,
I test whether gender has a statistically significant impact on the likelihood of choosing
to edit biographies of women versus biographies of men, even when controlling for other
characteristics of the respondent. Throughout this part, the sample is restricted to the control
group in the survey experiment.
Table 7 presents the likelihood of choosing to edit a Wikipedia article about a woman
over an article about a man, by the gender of the survey respondent. The first row shows
that 31 percent of female respondents choose to edit a biography of a woman, and only 11
percent of male respondents do that. Note that 31 percent of female respondents choosing to
edit a biography of a woman is not a small percentage, because they had to choose from the
existing biographies in Wikipedia and only 15.2 percent of those are biographies of women.
Rows 2 - 7 of Table 7 present the statistic by survey respondents’ education, age, and
whether they have edited Wikipedia in the past. In all the groups, female respondents are
more likely to edit biographies of women. The last rows of the table restrict attention to
subgroups of biographies (as defined in Table C.1): first, lawyers, politicians, scientists, etc.
(group Professionals); and second, writers, painters, composers, etc. (group Culture). A large
gender difference remains in both groups.
Table 8 predicts whether survey respondents chose to edit a Wikipedia article about a
women based on respondents’ gender and other characteristics. The characteristics included
are an indicator for being at least 35 years old, having a college degree, being unemployed,
working at least 35 hours per week, using Wikipedia daily, using Wikipedia at least weekly,
and having edited Wikipedia in the past. The results show that women are statistically
significantly more likely than men to edit an article about a women versus a man; no other
included characteristic has a significant impact.
3.2.3
Is the allocation of editorial input proportional to readership?
This section analyzes whether the allocation of editorial input to the Wikipedia articles about
men and women is proportional to readership. Each day, there is a large share of Wikipedia
articles that noone reads. I will start by restricting attention to those. Second, I will look at
all the biographies. The data on readership is from page views for the month of September
2014 for more than 900,000 biographies in Wikipedia.
First, I simply document that on a typical day, the percentage of biographies of men that
no one reads is larger than that of women. Table 9 presents the daily statistics regarding
the percentage of biographies in Wikipedia that no one reads. On a typical (median) day in
15
September 2014, no one read 26 percent of the biographies of men versus only 16 percent of
the biographies of women.26 Note that over a longer period of time almost all the articles
have readers. The above statistics could be interpreted that at the lower tail of the readership
distribution, biographies of men have fewer readers than biographies of women. We see this
also when we look at the whole distribution of readership.
Figure 3 presents the cumulative distribution of the number of page views of biographies
of men and women, where the page views are divided by the number of editors in the sample
who edited the page. As confirmed by the Kolmogorov-Smirnov test in Table C.3 in Appendix
C, biographies of women receive more page views per editor. Appendix B.1 repeats the
analysis for survey data, and results are similar. Overall, this seems to imply that demand
side cannot be the only reason for gender inequality in Wikipedia coverage. Instead, gender
inequality in coverage seems to originate, at least partially, from the supply side of Wikipedia
production.
3.3
How to reduce gender inequality in Wikipedia?
There are many potential ways how to reduce gender inequality in Wikipedia. I look at only
one of these, which is providing information. The main reason to concentrate on this one, is
that it is cheap and simple to implement.
This section presents evidence from a randomized survey experiment on the impact
of providing information about gender inequality in Wikipedia. First, I look at how the
treatment affected respondents’ answers regarding their choice of which Wikipedia article to
edit. Specifically, I’m interested in the gender of the person whose biography they chose to
edit. Second, I look at whether the treatment affects answers about the likelihood of editing
Wikipedia in the future.
3.3.1
Choice of pages to edit
Table 10 presents the effect of the treatment on the answers about the choice of the Wikipedia
article respondents would edit. Column 1 shows that the treatment is associated with a 95
percent increase in the share of women’s biographies. Note that in the control group, the
share of women’s biographies is only 20 percent. With the treatment, it gets close to 40
percent. Inclusion of demographic characteristics (column 2) decreases the scaled treatment
effect from 95 to 90 percent.
The scaled treatment effect is larger for men than women as shown in columns 3-6.
26
The median is taken over the days, and it is calculated separately for the biographies of men and the
biographies of women.
16
Column 3 shows that the treatment in the case of men is associated with a 119 percent
increase in the share of women’s biographies. The result is robust to inclusion of demographic
characteristics (column 4). Column 5 shows that the treatment in the case of women is
associated with a 77 percent increase in the share of women’s biographies, and it is robust to
inclusion of demographic characteristics (column 6). The larger treatment effect for male
respondents reflects the fact that the starting point for men was lower. In the control group,
only 11 percent of male respondents would edit biographies of women compared to 31 percent
of female respondents.
Overall, the treatment redirects editorial input and almost doubles the likelihood of
contributing to an article about a woman. Appendix B.2 describes how the treatment affects
the allocation of editorial input in terms of readership. There is no evidence that the treatment
makes the allocation significantly worse or better in terms of readership.
3.3.2
Participation
Table 11 presents the effect of the treatment on the answers about the likelihood of editing
Wikipedia in the future. Column 1 shows that the treatment is associated with a 35 percent
decrease in the likelihood of editing Wikipedia in the future. When including demographic
characteristics (column 2), the decrease is 37 percent. That the treatment decreases the
likelihood of editing is somewhat unexpected. Columns 3-6 look at the effect of the treatment
on men and women separately, and this helps to clarify the finding. We see that the treatment
is associated with a decrease in the likelihood of editing in the case of men (columns 3-4), and
it has no effect on women (columns 5-6). Namely, column 3 shows that the treatment in the
case of men is associated with a 62 percent decrease in editing. When including demographic
characteristics (column 4), the decrease is 64 percent. To analyze the robustness of the
finding regarding the likelihood of editing, I restrict the sample using three characteristics:
Wikipedia use, beliefs about competence, and writing user reviews. The results from the
restricted samples are presented in Table C.4 in Appendix C and are similar to those from
the main sample.
The information that the majority of Wikipedia editors are men, leads men to reduce
their editing effort, but it does not change the behavior of women. There are other examples
in the literature where informational treatment has backfired (e.g., Beshears, Choi, Laibson,
Madrian, and Milkman, 2015). On the one hand, the informational treatment in this paper
can be considered a moral reminder and, on the other hand, a peer information intervention
(information about what peers typically do, also called social norms marketing). There is
mixed evidence on whether moral reminders work.27 For example, Pruckner and Sausgruber
27
For an overview of the impact of moral reminders on tax compliance, see Luttmer and Singhal (2014).
17
(2013) find that a moral reminder increases honesty in payments, but Fellner, Sausgruber,
and Traxler (2013) find that moral reminders have no effect on tax compliance. Note that in
the treatment in this paper, it’s less costly for men to react to the moral reminder than for
women. Men can reduce the gender gap among editors by editing less, and this is easy to
do in terms of effort; but for women to reduce gender inequality, the solution is more costly,
they should edit more. That women don’t report an increase in editing is in line with the
studies that have found that peer information intervention moves behavior closer to the peer
group average.28 There is also some evidence from field experiments that information about
negative social norms does not encourage good behavior (e.g., Yeomans and Herberich, 2014).
4
Conclusions
This is the first comprehensive study of gender inequality in a new media environment such as
Wikipedia. Combining survey data, experimental evidence, and a large dataset of Wikipedia
editing histories, I study the reasons why women are less likely to contribute to Wikipedia, the
implications of this gender gap, and what can be done about it. The findings are important
for a widely used reference source like Wikipedia and there are wider implications for both
traditional and new media.29
I find that men and women tend to contribute to different topics. This implies that
gender gap among contributors leads to an unequal coverage of topics. Moreover, I find
that almost half of the gender gap in contributions is explained by gender differences in two
characteristics: frequency of Wikipedia use and belief about one’s competence. Combining
the difference in Wikipedia use with the finding that men and women contribute to different
topics, it points toward a possible equilibrium effect: if a media channel covers fewer topics
that women are interested in, then women use it less frequently and are less interested in
contributing.
The results of the paper suggest that providing information to the editors about the
gender inequality can alleviate the gender gap in contributions. The informational treatment
changes the allocation of editorial input. Although, it comes at the cost of decreasing the
editorial input of men. The result provides an example where encouraging gender equality
can partially backfire. Wikipedia has set a goal to increase the share of female editors. One
28
For example, information about peers moved outcomes closer to the average in case of energy conservation
(Schultz, Nolan, Cialdini, Goldstein, and Griskevicius, 2007), charitable giving (Croson and Shang, 2008;
Jones and Linardi, 2014), restaurant dining choices (Cai, Chen, and Fang, 2009), and contributions of movie
ratings (Chen, Harper, Konstan, and Li, 2010).
29
Wikies and other online collaboration tools are used in many public and private content production
environments. For example, a for-profit site, Wikia hosts several hundred thousand wikies on a variety of
topics.
18
way to achieve this is by discouraging male editors. However, this might not be desirable,
especially, considering the finding by Aaltonen and Seiler (2015) that past contributions tend
to motivate further contributions in Wikipedia. The implication for Wikipedia and other
forms of media is that it is important to balance the efforts of attracting new contributors
and keeping the current ones.
This paper has concentrated on gender, but more broadly, it highlights the importance of
representing diverse opinions. Similar issues arise in case of other demographic characteristics.
For example, it has been noted that information about black history is lacking in Wikipedia
and the Wikimedia Foundation is trying to engage new editors in order to change that.30
References
Aaltonen, A., and S. Seiler (2015): “Cumulative Growth in User Generated Content
Production: Evidence from Wikipedia,” Management Science, p. forthcoming.
Aguiar, M., and E. Hurst (2007): “Measuring Trends in Leisure: The Allocation of Time
Over Five Decades,” The Quarterly Journal of Economics, 122(3), 969–1006.
Algan, Y., Y. Benkler, M. F. Morell, and J. Hergueux (2013): “Cooperation in
a Peer Production Economy: Experimental Evidence from Wikipedia,” Working Paper.
NBER Workshop on the Economics of IT and Digitization.
Beaman, L., R. Chattopadhyay, E. Duflo, R. Pande, and P. Topalova (2009):
“Powerful Women: Does Exposure Reduce Bias?,” The Quarterly Journal of Economics,
124(4), 1497–1540.
Beaman, L., E. Duflo, R. Pande, and P. Topalova (2012): “Female Leadership Raises
Aspirations and Educational Attainment for Girls: A Policy Experiment in India,” Science,
335(6068), 582–586.
Bertrand, M. (2011): “New Perspectives on Gender,” in Handbook of Labor Economics,
ed. by David Card and Orley Ashenfelter, vol. Volume 4, Part B, pp. 1543–1590. Elsevier.
Beshears, J., J. J. Choi, D. Laibson, B. C. Madrian, and K. L. Milkman (2015):
“The Effect of Providing Peer Information on Retirement Savings Decisions,” The Journal
of Finance, p. forthcoming.
30
Smith, Jada F. 2015.
“Howard University Fills in Wikipedias Gaps in Black
History,”
New York Times,
February 19.
http://www.nytimes.com/2015/02/20/us/
at-howard-a-historically-black-university-filling-in-wikipedias-gaps-in-color.html.
19
Cai, H., Y. Chen, and H. Fang (2009): “Observational Learning: Evidence from a
Randomized Natural Field Experiment,” American Economic Review, 99(3), 864–82.
Chen, Y., F. M. Harper, J. Konstan, and S. X. Li (2010): “Social Comparisons and
Contributions to Online Communities: A Field Experiment on MovieLens,” American
Economic Review, 100(4), 1358–98.
Chiang, C.-F., and B. Knight (2011): “Media Bias and Influence: Evidence from
Newspaper Endorsements,” The Review of Economic Studies, 78(3), 795–820.
Coffman, K. B. (2014): “Evidence on Self-Stereotyping and the Contribution of Ideas,”
The Quarterly Journal of Economics, 129(4), 1625–1660.
Croson, R., and U. Gneezy (2009): “Gender Differences in Preferences,” Journal of
Economic Literature, 47(2), 448–474.
Croson, R., and J. Y. Shang (2008): “The impact of downward social information on
contribution decisions,” Experimental Economics, 11(3), 221–233.
DellaVigna, S., R. Enikolopov, V. Mironova, M. Petrova, and E. Zhuravskaya
(2014): “Cross-Border Media and Nationalism: Evidence from Serbian Radio in Croatia,”
American Economic Journal: Applied Economics, 6(3), 103–132.
Eckel, C. C., and P. J. Grossman (2008): “Differences in the Economic Decisions of Men
and Women: Experimental Evidence,” in Handbook of Experimental Economics Results,
ed. by Charles R. Plott and Vernon L. Smith, vol. Volume 1, pp. 509–519. Elsevier.
Enikolopov, R., M. Petrova, and E. Zhuravskaya (2011): “Media and Political
Persuasion: Evidence from Russia,” American Economic Review, 101(7), 3253–3285.
Fairlie, R. W. (2005): “An extension of the Blinder-Oaxaca decomposition technique to
logit and probit models,” Journal of Economic and Social Measurement, 30(4), 305–316.
Fairlie, R. W., F. Hoffmann, and P. Oreopoulos (2014): “A Community College
Instructor Like Me: Race and Ethnicity Interactions in the Classroom,” American Economic
Review, 104(8), 2567–2591.
Fellner, G., R. Sausgruber, and C. Traxler (2013): “Testing Enforcement Strategies
in the Field: Threat, Moral Appeal and Social Information,” Journal of the European
Economic Association, 11(3), 634–660.
20
Ferrara, E. L., A. Chong, and S. Duryea (2012): “Soap Operas and Fertility: Evidence
from Brazil,” American Economic Journal: Applied Economics, 4(4), 1–31.
Gentzkow, M., and J. M. Shapiro (2010): “What Drives Media Slant? Evidence From
U.S. Daily Newspapers,” Econometrica, 78(1), 35–71.
Ghose, A., and S. P. Han (2011): “An Empirical Analysis of User Content Generation
and Usage Behavior on the Mobile Internet,” Management Science, 57(9), 1671–1691.
Greenstein, S., and F. Zhu (2012): “Is Wikipedia Biased?,” American Economic Review:
Papers and Proceedings, 102(3), 343–348.
(2014): “Do Experts or Collective Intelligence Write with More Bias? Evidence
from Encyclopedia Britannica and Wikipedia,” Harvard Business School Working Paper,
15-023.
Groseclose, T., and J. Milyo (2005): “A Measure of Media Bias,” The Quarterly Journal
of Economics, 120(4), 1191–1237.
Horton, J. J., D. G. Rand, and R. J. Zeckhauser (2011): “The online laboratory:
conducting experiments in a real labor market,” Experimental Economics, 14(3), 399–425.
Jensen, R., and E. Oster (2009): “The Power of TV: Cable Television and Women’s
Status in India,” The Quarterly Journal of Economics, 124(3), 1057–1094.
Jones, D., and S. Linardi (2014): “Wallflowers: Experimental Evidence of an Aversion to
Standing Out,” Management Science, 60(7), 1757–1771.
Kuziemko, I., M. I. Norton, E. Saez, and S. Stantcheva (2015): “How Elastic
Are Preferences for Redistribution? Evidence from Randomized Survey Experiments,”
American Economic Review, 105(4), 1478–1508.
Luttmer, E. F. P., and M. Singhal (2014): “Tax Morale,” Journal of Economic
Perspectives, 28(4), 149–68.
Mason, W., and S. Suri (2012): “Conducting behavioral research on Amazon’s Mechanical
Turk,” Behavior Research Methods, 44(1), 1–23.
Nagaraj, A. (2014): “Does Copyright Affect Creative Reuse? Evidence from the Digitization
of Baseball Digest,” manuscript.
Niederle, M., and L. Vesterlund (2011): “Gender and Competition,” Annual Review
of Economics, 3(1), 601–630.
21
Oaxaca, R. (1973): “Male-Female Wage Differentials in Urban Labor Markets,” International Economic Review, 14(3), 693.
Oberholzer-Gee, F., and J. Waldfogel (2009): “Media Markets and Localism: Does
Local News en Espanol Boost Hispanic Voter Turnout?,” American Economic Review,
99(5), 2120–2128.
Pruckner, G. J., and R. Sausgruber (2013): “Honesty on the Streets: A Field Study on
Newspaper Purchasing,” Journal of the European Economic Association, 11(3), 661–679.
Reagle, J., and L. Rhue (2011): “Gender Bias in Wikipedia and Britannica,” International
Journal of Communication, 5(0), 21.
Reuter, J., and E. Zitzewitz (2006): “Do Ads Influence Editors? Advertising and Bias
in the Financial Media,” The Quarterly Journal of Economics, 121(1), 197–227.
Schultz, P. W., J. M. Nolan, R. B. Cialdini, N. J. Goldstein, and V. Griskevicius
(2007): “The Constructive, Destructive, and Reconstructive Power of Social Norms,”
Psychological Science, 18(5), 429–434.
Shriver, S. K., H. S. Nair, and R. Hofstetter (2013): “Social Ties and UserGenerated Content: Evidence from an Online Social Network,” Management Science, 59(6),
1425–1443.
Sun, M., and F. Zhu (2013): “Ad Revenue and Content Commercialization: Evidence from
Blogs,” Management Science, 59(10), 2314–2331.
Tella, R. D., and I. Franceschelli (2011): “Government Advertising and Media
Coverage of Corruption Scandals,” American Economic Journal: Applied Economics, 3(4),
119–151.
Toubia, O., and A. T. Stephen (2013): “Intrinsic vs. Image-Related Utility in Social
Media: Why Do People Contribute Content to Twitter?,” Marketing Science, 32(3),
368–392.
Wolbrecht, C., and D. E. Campbell (2007): “Leading by Example: Female Members
of Parliament as Political Role Models,” American Journal of Political Science, 51(4),
921–939.
Yeomans, M., and D. Herberich (2014): “An experimental test of the effect of negative social norms on energy-efficient investments,” Journal of Economic Behavior &
Organization, 108, 187–197.
22
Yildirim, P., E. Gal-Or, and T. Geylani (2013): “User-Generated Content and Bias
in News Media,” Management Science, 59(12), 2655–2666.
Zhang, X., and F. Zhu (2011): “Group Size and Incentives to Contribute: A Natural
Experiment at Chinese Wikipedia,” The American Economic Review, 101(4), 1601–1615.
23
Tables
Table 1: Summary statistics of Wikipedia editors, split by gender
Male editors
Female editors
Percentiles
25th 50th 75th 25th 50th 75th
Number of biographies edited, per year
1
2
4
1
1
3
Length of text added, per year
22
169 1169 26
188 1247
Length of text added, per biography
13
80
294
16
98
474
Observations
65398
9419
Note: A unit of observation is an editor. Columns 1-3, present the 25th, 50th, and 75th percentile
for male editors, and columns 4 - 6, for female editors. Length of text added measures in characters
the length of text added to biographies in Wikipedia. Sample of editors includes all the editors who
reported their gender and who have edited the English language Wikipedia articles about human
beings in the Wikidata final sample.
Table 2: Mean values of survey respondents’ characteristics
All respondents Only editors
Female
0.495
0.315
At least 35 years old
0.354
0.286
Has college degree
0.540
0.596
Works at least 35 hours per week
0.577
0.592
Unemployed
0.100
0.085
Use Wikipedia daily
0.277
0.460
Use Wikipedia at least weekly
0.811
0.906
Has edited Wikipedia
0.219
1.000
Observations
974
213
Note: Each cell presents the mean value of a characteristic that takes values either 1 or 0. In column
1, the sample includes all valid survey responses. In column 2, the sample is restricted to those who
have edited Wikipedia in the past.
24
Table 3: Summary statistics of survey respondents in the control group, split by gender
Men Women
p-value
Demographic characteristics
0.320 0.382
0.154
Has college degree
0.514 0.536
0.611
0.653 0.459
0.000
Unemployed
0.089 0.116
0.322
Wikipedia
Use Wikipedia daily
0.359 0.176
0.000
0.838 0.742
0.009
0.313 0.163
0.000
Likely to edit Wikipedia
0.147 0.094
0.077
Believe: more competent than other editors
0.085 0.043
0.059
Believe: at least as competent as other editors 0.761 0.670
0.025
Other
Donated to charity
0.475 0.554
0.081
Volunteering at least one hour per week
0.332 0.365
0.447
Wrote user review in previous six months
0.517 0.545
0.540
Leisure at least 3 hours per weekday
0.741 0.734
0.852
Observations
259
233
Note: The sample is restricted to the control group. Columns 1 and 2, present the means of the
survey respondents characteristics separately for men and women. Column 3, presents the p-value
of the t-test for whether the difference between men and women is significantly different from zero.
Likely to edit Wikipedia is an indicator variable that takes value one if answered “very likely” or
“quite likely” to question “How likely or unlikely is it that you will actually edit Wikipedia over the
next 30 days?” Believe: more competent than other editors is an indicator variable that takes value 1
if answered “more knowledgeable/competent” to the question “Think of the Wikipedia article of the
human being that you chose. On the whole, do you think that you are more or less knowledgeable
and competent to edit the article than other people who will edit it in the future?” Believe: at least
as competent as other editors is an indicator variable that takes value 1 if answered either “more
knowledgeable/competent”, “somewhat more”, or “same” to the same question above.
25
Table 4: Estimated role of survey respondents’ characteristics in Wikipedia editing
(1)
(2)
Female
-0.669***
-0.148
(0.245)
(0.317)
-0.104
-0.184
(0.251)
(0.322)
Has college degree
0.429*
0.310
(0.238)
(0.310)
Unemployed
-0.055
0.105
(0.411)
(0.597)
-0.345
0.192
(0.256)
(0.338)
Use Wikipedia daily
1.025*** 1.156***
(0.254)
(0.331)
0.475
0.063
(0.358)
(0.468)
0.782*
0.414
(0.415)
(0.479)
Believe: at least as competent as other editors 0.743** 1.946***
(0.297)
(0.614)
Donated to charity
0.238
0.387
(0.253)
(0.336)
-0.146
0.345
(0.256)
(0.322)
0.552**
0.294
(0.236)
(0.306)
-0.025
0.094
(0.260)
(0.338)
Observations
492
492
Log-likelihood
-238.921 -156.098
Note: Each column presents estimation results from a separate logit equation. In column 1, the
dependent variable is an indicator for whether the individual has edited Wikipedia in the past. In
column 2, the dependent variable is whether the individual reports he is likely to edit Wikipedia
in the next 30 days, specifically, the variable takes value one if the respondent answered that he
is either “very likely” or “quite likely” to the question: “How likely or unlikely is it that you will
actually edit Wikipedia over the next 30 days?” Believe: more competent than other editors is an
indicator variable that takes value 1 if answered “More knowledgeable/competent” to the question
“Think of the Wikipedia article of the human being that you chose. On the whole, do you think that
you are more or less knowledgeable and competent to edit the article than other people who will
edit it in the future?” Believe: at least as competent as other editors is an indicator variable that
takes value 1 if answered either “more knowledgeable/competent”, “somewhat more”, or “same”
to the same question above. The sample is restricted to the control group. Standard errors are in
parantheses. *** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant
at the 10 percent level.
26
Table 5: Decomposition of gender gap in Wikipedia editing
Total gap Explained gap Explained percentage
Dependent variable: Has edited
0.1497
0.0642
42.9
Dependent variable: Likely to edit
0.0523
0.0412
78.8
Note: Each row presents a separate decomposition. In the first row, the dependent variable is an
indicator for having editing Wikipedia in the past. In the second row, the dependent variable is
an indicator for whether the individual believes he is likely to edit Wikipedia in the next 30 days,
defined in the same way as in Table 3. There are four explanatory variables: indicator for daily
Wikipedia use, indicator for at least weekly Wikipedia use, Believe: more competent than other
editors, and Believe: at least as competent as other editors, defined in the same way as in Table
4. In the columns, Total gap is the difference between the average outcome variable of men and
women; Explained gap is the difference between the outcome variable of men and women that is
explained by the differences in these explanatory variables; Explained percentage is the percentage
of the Total gap that is explained by the differences in these explanatory variables. The sample
consists of the survey respondents in the control group.
Table 6: The average (over editors) percentage of contributions to the biographies of women
vs men, by the gender of Wikipedia editor
Mean % of biographies of women p-value
Male editors
Female editors
Total length of text added
16.6
31.4
0.000
– Subset of biographies: Professionals
8.0
16.7
0.000
– Subset of biographies: Culture
15.7
31.7
0.000
Number of biographies edited
17.1
31.1
0.000
– Subset of biographies: Professionals
8.3
15.9
0.000
– Subset of biographies: Culture
15.7
31.5
0.000
Note: A unit of observation is a Wikipedia editor. Columns 1 and 2, present the means separetely
for male and female Wikipedia editors. Column 3, presents the p-value of the t-test for whether the
difference between male and female editors is significantly different from zero. In rows 1 - 3, the
average percentage of contributions to the biographies of women is measured in terms of the total
length of text added. In rows 4 - 6, it is measured in terms of the number of biographies. Total
sample includes 65,398 male and 9,419 female editors. In rows 2 and 5, the sample is restricted to
the biographies of lawyers, politicians, scientists, etc (Professionals). In rows 3 and 6, the sample is
restricted to the biographies of writers, painters, composers, etc (Culture). The exact definition of
these groups of professions is in Table C.1.
27
Table 7: Likelihood of choosing to edit a Wikipedia article about a woman vs a man, by the
gender of survey respondent
Likelihood of choosing an article about a woman
Male respondents
Female respondents
Total
0.112
0.305
Demographic characteristics
0.148
0.395
Hasn’t edited Wikipedia
0.096
0.287
College
0.128
0.272
No college
0.095
0.343
Less than 35 years of age
0.084
0.270
At least 35 years of age
0.125
0.326
Wikipedia articles
Subset of articles: Professionals
0.076
0.205
Subset of articles: Culture
0.123
0.279
p-value
0.000
0.003
0.000
0.004
0.000
0.001
0.000
0.015
0.016
Note: A unit of observation is a survey respondent. Each survey respondent chose a Wikipedia
article either about a man or a woman. Columns 1 and 2 present the mean of the indicator variable
that the chosen article was about a woman, separately for male and female survey respondents.
Column 3, presents the p-value of the t-test for whether the difference between male and female
respondents is significantly different from zero. In the first row, the sample consists of the survey
respondents in the control group, there are 259 male and 233 female respondents. In the following
rows, the sample is further restricted by the survey respondents’ demographic characteristics that
are listed in the leftmost column. In the last two rows, the sample is restricted to the subgroups of
biographies: group Professionals includes lawyers, politicians, scientists, etc; and group Culture
includes writers, painters, composers, etc; the exact definition of groups is in Table C.1.
28
Table 8: Estimated role of survey respondents’ characteristics on the likelihood of choosing
to edit a Wikipedia article about a woman vs a man
Article about a woman
Female
1.368***
(0.261)
-0.308
(0.253)
Has college degree
-0.109
(0.239)
Unemployed
-0.325
(0.415)
-0.165
(0.252)
Use Wikipedia daily
0.138
(0.292)
0.055
(0.305)
0.460
(0.282)
Observations
492
Log-likelihood
-230.888
Note: Estimation results are from a logit equation. Dependent variable is an indicator for whether the
respondent chose to edit an article about a woman. The explanatory variables are the characteristics
of the survey respondent. The sample is restricted to the control group. Standard errors are in
parantheses. *** Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant
at the 10 percent level.
Table 9: The daily percentage of biographies
Median
Biographies of men
25.8
Biographies of women
16.2
in Wikipedia
Min Max
18.6 28.4
11.9 18.6
that noone read
Obs
30
30
Note: A unit of observation is a day. The sample consists of daily data of page views from September
2014 for 784,178 biographies of men and 140,193 biographies of women in Wikipedia.
29
Table 10: The effect of treatment on the likelihood of choosing to edit a Wikipedia article
about a woman
Total
Male respondents Female respondents
(1)
(2)
(3)
(4)
(5)
(6)
Treated
0.193*** 0.182*** 0.133*** 0.132*** 0.233*** 0.236***
(0.029)
(0.028)
(0.034)
(0.034)
(0.044)
(0.044)
Covariates
No
Yes
No
Yes
No
Yes
Control group mean
0.203
0.203
0.112
0.112
0.305
0.305
Scaled treatment effect
0.950
0.898
1.185
1.182
0.766
0.774
Observations
974
974
492
492
482
482
Note: Each columns presents estimates from a separate regression. Dependent variable is an
indicator for whether the respondent chose to edit an article about a woman. Covariates include
demographic characteristics from Table 3. The sample includes both the control and treatment group
of the survey respondents. The sample is restricted to men in columns 3 - 4 and women in columns
5-6. In columns 2, 4, and 6, regression include the standard set of demographic characteristics used
in the paper and described below Table 4. Scaled treatment effect measures the percentage change
of the share of women’s pages. Standard errors are in parantheses. *** Significant at the 1 percent
level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
Table 11: The effect of treatment on the likelihood of editing Wikipedia
Treated
Covariates
Control group mean
Scaled treatment effect
Observations
Total
(1)
(2)
-0.043** -0.045**
(0.019)
(0.019)
No
Yes
0.122
0.122
-0.354
-0.369
974
974
Male respondents
Female respondents
(3)
(4)
(5)
(6)
-0.091*** -0.094*** 0.006
0.002
(0.027)
(0.027)
(0.027)
(0.027)
No
Yes
No
Yes
0.147
0.147
0.094
0.094
-0.620
-0.638
0.063
0.023
492
492
482
482
indicator for whether the individual believes he is likely to edit Wikipedia in the next 30 days,
defined in the same way as in Table 3. Covariates include demographic characteristics from Table 3.
The sample includes both the control and treatment group of the survey respondents. The sample
is restricted to men in columns 3 - 4 and women in columns 5-6. In columns 2, 4, and 6, regression
include the standard set of demographic characteristics used in the paper and described below Table
4. Scaled treatment effect measures the percentage change of the share of respondents who believe
it is likely they will edit Wikipedia. Standard errors are in parantheses. *** Significant at the 1
percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
30
12
0
0
2
4
6
8
10
Percentage of female editors
Number of active editors, thousands
10
20
30
Figures
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
Year
Number of active editors
Perc. of female editors
Figure 1: Number of Wikipedia editors (left y-axes) and percentage of female editors (right y-axes),
in 2002 - 2013
Note: The dataset is restricted to the registered editors of Wikipedia who have reported their gender
and edited at least one article in a given calendar year.
31
Percentage of biographies of women
10
20
30
0
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
Year
Male editors
Female editors
Figure 2: Average (over Wikipedia editors) percentage of contributions to the biographies of
women vs men, in 2002 - 2013
Note: The percentage is calculated as the average (over Wikipedia editors) percentage of the total
length of text added to the biographies of women vs men.
1
Cumulative Probability
.8
.6
.4
.2
0
0
200
400
600
800
Number of page views divided by number of editors per biography
c.d.f.: biographies of men
1000
c.d.f.: biographies of women
Figure 3: Cumulative distribution functions of the number of page views divided by the
number of editors per biography
Note: The sample includes page views data from September 2014 for 784,178 biographies of men
and 140,193 biographies of women in Wikipedia. The number of page views divided by the number
of editors is constructed based on the pages and editors in the sample of historical editing data from
Wikipedia.
32
A
Appendix: Survey attrition
The survey had an overall attrition rate of 19.4%.31 This section tests whether attrition is
random or depends on demographic characteristics and treatment status.
98% of the individuals who dropped out, did that on the page of the survey that displayed
the task of choosing the Wikipedia page which to edit and finding information what to add to
the page. This is probably where the participants realized that the task is more difficult than
they expected. Two percent of those who dropped out, did that before seeing the treatment
page. None dropped out of on the treatment page. Altogether, this is likely to indicate that
treatment itself was not a reason for dropping out.
Table C.5 in Appendix C shows that although, attrition is not random, it doesn’t depend on
treatment status nor gender. However, as could be expected, people who use or edit Wikipedia,
are more likely to finish the survey. Table C.5 presents estimates from linear regressions32
of the form: F inishedSurveyi = Constant + βV ariablei + εi , where the dependent variable
F inishedSurveyi is an indicator of whether person i finished the survey. Each row of the table
presents the estimate of coefficient β and corresponding p-value from a separate regression
where the V ariablei of the regression is listed in the first column. In each regression, the
sample includes only those respondents who have remained in the survey to answer the
question; in case of treatment status, the sample includes those who have remained in the
survey until they were assigned the status. The first row of the table shows that gender has
no statistically significant impact on finishing the survey. The last row shows that treatment
status has no statistically significant impact either. The table also shows that attrition
depends on demographic characteristics and those who use Wikipedia more are more likely
to finish the survey. Specifically, people who have at least a college degree are more likely to
finish the survey; and those who are older or unemployed are less likely to finish the survey.
People who use Wikipedia at least weekly or who have edited it, are more likely to finish the
survey.
Overall the sample is not fully representative of the general population, although, it is
closer to it than students in a class room. The sample is more likely to be younger and spend
more time online than the general population. In terms of these characteristics, it is the
relevant sample to survey, as these are the people who are more likely to potentially edit
Wikipedia. Similarly, in terms of attrition, the participants who are more likely to finish the
survey are the ones who use or edit Wikipedia. When interpreting the survey results, we
31
When counting the number of people who dropped out, I exclude 2 responses that don’t have unique
Amazon Mechanical Turk IDs. When calculating the attrition rate, I include only the 974 valid completed
responses and the uncompleted responses.
32
Estimates from analogous logit regressions are available from the author. The results are very similar.
33
should remember that these don’t represent the general population, but instead the group of
people for whom Wikipedia is more important.
To test it further that the attrition does not depend on treatment status, Table C.6 in
Appendix C presents results from a regression, where the dependent variable is an indicator
for finishing the survey. It is regressed on the treatment status, demographic characteristics
and treatment status interacted with the demographic characteristics. Treatment status has
no significant effect on finishing the survey. There are two characteristics that matter for
attrition: those who have edited Wikipedia before and those who are younger are more likely
to finish the survey.
Table C.7 in Appendix C shows that conditional on finishing the survey, assignment into
treatment group was random in terms of most characteristics. The table presents estimates
from linear regressions33 of the form: T reatmentGroupi = Constant + βV ariablei + εi ,
where the dependent variable T reatmentGroupi equals one if person i was assigned into the
treatment group and zero if he was assigned into the control group. Each row of the table
presents the estimate of coefficient β and corresponding p-value from a separate regression
where the V ariablei of the regression is listed in the first column. In each regression, the
sample includes only those respondents who finished the survey. The table shows that
conditional on finishing the survey, assignment into treatment group was random in terms of
the demographic characteristics, attitudes towards Wikipedia, frequency of Wikipedia use,
charity, writing user reviews, and free time. From the 15 outcomes, two are significant at the
10 percent level. Those who have edited Wikipedia before, conditional on finishing the survey
are less likely to be in the treatment group, with p-value equal 0.077. Note that the question
was asked before treatment, hence, the treatment couldn’t affect the answers. Those who
volunteer more are also less likely to be in treatment group, with p-value 0.075.
The treatment provided information about Wikipedia and was designed to change
Wikipedia editing behavior. Two concerns were that the treatment could lead to differential attrition or change the answers to the survey questions. From Table C.5 we saw that
there is no evidence that the treatment lead to differential attrition. Table C.7 showed that
in terms of most characteristics the randomization into treatment was successful. In any case,
when analyzing the survey responses, I restrict attention to the control group, to make sure
that the treatment does not affect the answers.
33
Estimates from analogous logit regressions are available from the author. The results are very similar.
34
B
B.1
Appendix: Additional empirical analysis
Is the editorial input proportional to readership: evidence
from the survey
The evidence from the Wikipedia editing histories showed that pages of men receive relatively
fewer readers than the pages of women. A similar exercise with the survey data confirms the
results.
Figure D.2 in Appendix D presents the cumulative distribution of the number of page
views of the Wikipedia pages of men and women, where the page views are divided by the
number of survey respondents editing the page. The measure is constructed only based on
the editors and pages from the control group in the survey. As partly confirmed by the
Kolmogorov-Smirnov test, presented in Table C.8 in Appendix C, the c.d.f. of the pages of
women seems to be larger than that of the pages of men. In Table C.8, the null hypothesis is
that the samples are drawn from the same distribution. The first row of the table presents
the test statistic and p-value from the one-sided hypothesis test, where alternative hypothesis
is that the c.d.f. of the pages of women is larger than the c.d.f. of the pages of men. The
one-sided test rejects the null hypothesis at the 10 percent significance level. The second row
of the table presents the test statistic and p-value from the two-sided hypothesis test, where
the alternative hypothesis is that the samples are drawn from two different distributions.
The two-sided test does not reject the null hypothesis.
B.2
In terms of total readership, does the treatment improve the
allocation of editorial input?
In this section, I compare the allocation of editorial input in the treatment and control group.
The goal is to learn whether in the treatment group compared to the control group more
editorial input is allocated to the pages with a larger or smaller number of readers.
Figure D.3 in Appendix D presents the cumulative distribution of the number of page
views of the Wikipedia pages in the control and treatment groups, where the page views are
divided by the number of survey respondents editing the page in control and treatment groups
respectively. As confirmed by the Kolmogorov-Smirnov test in Table C.9 in Appendix C, the
distributions look similar. In Table C.9, the null hypothesis is that the the samples are drawn
from the same distribution. The first row of the table presents the test statistic and p-value
from the one-sided hypothesis test, where alternative hypothesis is that the c.d.f. of the
pages in control group is smaller than the c.d.f. of the pages in treatment group. The second
row presents the results from the one-sided hypothesis test, where the alternative hypothesis
35
is that the c.d.f. of the pages in treatment group is smaller than the c.d.f. of the pages
in control group. The third row presents results from the two-sided hypothesis test, where
the alternative hypothesis is that the samples are drawn from two different distributions.
All three tests cannot reject the null hypothesis that the samples are drawn from the same
distribution at the 10 percent significance level.
Hence, we can conclude that although the treatment moved the allocation of editorial
input towards the pages of women, this didn’t make the allocation worse nor better in terms
of the readership.
36
C
Appendix: Additional tables
Table C.1: Professions of human beings in Wikidata
Profession
Number of professions
football player
105871
actor
42096
politician
40084
sportsperson
23040
baseball player
20423
writer
15515
cricketer
15448
lawyer
14812
painter
12746
ice hockey player
11751
composer
11692
priest
11242
film director
9794
basketball player
8561
singer
8122
journalist
7619
poet
7273
screenwriter
6807
musician
6325
diplomat
6214
Australian-rules footballer
6122
judge
5767
officer
5509
mathematician
5352
linguist
5229
photographer
4750
tennis player
4498
conductor
4428
boxer
4305
rugby league player
4222
physicist
4197
film producer
4192
economist
3922
author
3485
architect
3425
anthropologist
3082
bicycle racer
3065
Table continues on next page
37
Category
Other
Other
Professionals
Other
Other
Culture
Other
Professionals
Culture
Other
Culture
Professionals
Other
Other
Other
Professionals
Culture
Other
Other
Professionals
Other
Professionals
Professionals
Professionals
Professionals
Culture
Other
Culture
Other
Other
Professionals
Other
Professionals
Culture
Professionals
Professionals
Other
Profession
Number
basketball coach
golfer
explorer
singer-songwriter
chemist
botanist
film actor
rugby union player
astronomer
television actor
motorcycle racer
television presenter
sculptor
historian
theologian
model
philosopher
biologist
computer scientist
physician
pianist
field hockey player
voice actor
entrepreneur
alpine skier
swimmer
librarian
soldier
civil engineer
badminton player
Table continued from previous page
38
of professions Category
3048
Other
2919
Other
2856
Professionals
2841
Other
2651
Professionals
2585
Professionals
2432
Other
2374
Other
2271
Professionals
2164
Other
1961
Other
1924
Other
1915
Culture
1904
Professionals
1672
Professionals
1606
Other
1594
Professionals
1580
Professionals
1546
Professionals
1515
Professionals
1457
Culture
1452
Other
1409
Other
1282
Professionals
1281
Other
1255
Other
1211
Professionals
1159
Professionals
1075
Professionals
1064
Other
Table C.2: Estimated role of survey respondent’s characteristics for Wikipedia editing, by
gender
Has college degree
Unemployed
Use Wikipedia daily
Believe: at least as competent as other editors
Donated to charity
Observations
Log-likelihood
(1)
-0.415
(0.326)
0.673**
(0.301)
-0.319
(0.595)
-0.502
(0.335)
0.901***
(0.312)
0.419
(0.487)
0.774
(0.494)
0.405
(0.378)
0.223
(0.324)
-0.116
(0.336)
0.472
(0.298)
0.016
(0.332)
259
-144.593
(2)
0.313
(0.403)
-0.010
(0.411)
0.160
(0.598)
-0.116
(0.409)
1.236***
(0.456)
0.492
(0.553)
0.908
(0.794)
1.249**
(0.528)
0.343
(0.450)
-0.246
(0.422)
0.814**
(0.412)
-0.254
(0.451)
233
-90.465
(3)
-0.465
(0.445)
0.478
(0.409)
-0.442
(1.134)
-0.418
(0.452)
1.212***
(0.426)
0.724
(0.818)
0.493
(0.553)
2.533**
(1.058)
0.740
(0.459)
0.078
(0.446)
0.428
(0.407)
-0.236
(0.424)
259
-87.920
(4)
0.063
(0.498)
0.039
(0.508)
0.557
(0.770)
0.847
(0.535)
1.025*
(0.562)
-0.217
(0.627)
-0.566
(1.144)
1.574**
(0.771)
0.128
(0.543)
0.608
(0.500)
0.157
(0.495)
0.716
(0.623)
233
-63.547
Note: Each column presents estimation results from a separate logit equation. In columns 1 and 2,
the dependent variable is an indicator for whether the individual has edited Wikipedia in the past.
In columns 3 and 4, the dependent variable is whether the individual is likely to edit Wikipedia in
the next 30 days, specifically, the variable takes value one if the respondent answered that he is either
“very likely” or “quite likely” to the question: “How likely or unlikely is it that you will actually
edit Wikipedia over the next 30 days?” The sample in columns 1 and 3 is restricted to men in the
control group; in columns 2 and 4, it’s women in the control group. Believe: more competent than
other editors is an indicator variable that takes value 1 if answered “More knowledgeable/competent”
to the question “Think of the Wikipedia article of the human being that you chose. On the whole,
do you think that you are more or less knowledgeable and competent to edit the article than other
people who will edit it in the future?” Believe: at least as competent as other editors is an indicator
variable that takes value 1 if answered either “More knowledgeable/competent”, “Somewhat more”,
or “Same” to the same question above. Standard errors are in parantheses. *** Significant at the 1
percent level. ** Significant at the 5 percent level. * Significant at the 10 percent level.
39
Table C.3: Kolmogorov-Smirnov test of whether the samples of the number of page views
divided by the number of editors of the pages of men and women are drawn from the same
distribution
Test statistic p-value
One-sided hypothesis test
0.1414
0.000
Two-sided hypothesis test
0.1414
0.000
Note: The number of page views divided by the number of editors is constructed based on the
pages and editors in the sample of historical editing data from Wikipedia. Page views data from
September 1, 2014. The null hypothesis is that the samples are drawn from the same distribution.
In row 1, the alternatie hypothesis is that the c.d.f. of the pages of women is larger than the c.d.f.
of the pages of men. In row 2, the alternative hypothesis is that the samples are drawn from two
different distributions.
Table C.4: Differences in the effect of treatment on the likelihood of editing Wikipedia: by
Wikipedia use, believed competence, and whether writes user reviews
Wikipedia use
Competence
Writing reviews
(1)
(2)
(3)
(4)
(5)
(6)
Treated
-0.048** -0.049** -0.064** -0.066*** -0.058** -0.059**
(0.022)
(0.022)
(0.025)
(0.026)
(0.028)
(0.028)
Covariates
No
Yes
No
Yes
No
Yes
Control group mean
0.136
0.136
0.161
0.161
0.146
0.146
Scaled treatment effect -0.356
-0.360
-0.399
-0.410
-0.400
-0.404
Observations
790
790
693
693
513
513
indicator for whether the individual believes he is likely to edit Wikipedia in the next 30 days,
defined in the same way as in Table 3. Covariates include demographic characteristics from Table 3.
The sample includes both the control and treatment group of the survey respondents. In columns
1-2, the sample is restricted to those who use Wikipedia at least weekly. In columns 3-4 the sample
is restricted to those who believe they are at least as competent and knowledgeable as other people
who will edit Wikipedia in the future. In columns 5-6 the sample is restricted to those who wrote a
user review during the past six months. In columns 2, 4, and 6, regression include the standard set of
demographic characteristics used in the paper and described below Table 4. Scaled treatment effect
measures the percentage change of the share of respondents who believe it is likely they will edit
Wikipedia. Standard errors are in parantheses. *** Significant at the 1 percent level. ** Significant
at the 5 percent level. * Significant at the 10 percent level.
40
Table C.5: Ability of covariates to predict whether respondents finish the survey
Variable
Coef. p-value
Female
-0.026 0.252
-0.058 0.013
Has college degree
0.045
0.047
Works at least 35 hours per week 0.036
0.113
Unemployed
-0.095 0.008
0.057
0.045
Know can edit Wikipedia
0.050
0.146
0.116
0.000
Treatment group
0.029
0.200
Note: Each row presents estimates from a separate linear regression of the form F inishedSurveyi =
Constant + βV ariablei + εi , where the V ariablei is listed in the first column. A unit of observation
is a survey respondent. In each regression, the sample includes only those respondents who have
remained in the survey to answer the question; in case of treatment status, the sample includes
those who have remained in the survey until they were assigned the status. Number of observations
is 1205 for demographic characteristics and 1204 for Wikipedia related questions and treatment
status.
41
Table C.6: Ability of covariates to predict whether respondents finish the survey
(1)
(2)
Female
-0.015
-0.179
(0.156)
(0.215)
-0.323** -0.444**
(0.152)
(0.208)
Has college degree
0.205
0.344*
(0.153)
(0.208)
0.032
-0.088
(0.169)
(0.232)
Unemployed
-0.512** -0.565*
(0.232)
(0.319)
Use Wikipedia daily
-0.075
-0.211
(0.185)
(0.254)
0.207
0.067
(0.188)
(0.256)
0.864*** 0.882***
(0.241)
(0.315)
Treatment group
0.204
-0.439
(0.150)
(0.469)
Treatment group * Female
0.345
(0.314)
Treatment group * At least 35 years old
0.276
(0.308)
Treatment group * Has college degree
-0.291
(0.308)
Treatment group * Works at least 35 hours per week
0.265
(0.341)
Treatment group * Unemployed
0.085
(0.467)
Treatment group * Use Wikipedia daily
0.289
(0.376)
Treatment group * Use Wikipedia at least weekly
0.342
(0.378)
Treatment group * Has edited Wikipedia
-0.025
(0.492)
Constant
1.153*** 1.438***
(0.239)
(0.314)
Observations
1204
1204
Log-likelihood
-569.392 -566.992
Note: Each column presents estimates from a separate logit regression. Dependent variable is
an indicator of whether respondent finished the survey. Standard errors are in parentheses. ***
Significant at the 1 percent level. ** Significant at the 5 percent level. * Significant at the 10 percent
level.
42
Table C.7: Ability of covariates to predict treatment status, conditional on finishing the
survey
Variable
Coef. p-value
Female
0.043
0.180
0.010
0.761
Has college degree
0.032
0.323
0.033
0.307
Unemployed
-0.011 0.830
Use Wikipedia daily
0.012
0.733
0.061
0.138
-0.069 0.077
-0.003 0.963
Believe: at least as competent as other editors -0.015 0.678
Donated to charity
-0.035 0.275
-0.020 0.560
-0.008 0.811
Leisure at least 3 hours in a weekday
0.015
0.691
Note: Each row presents estimates from a separate regression of the form T reatmentGroupi =
Constant + βV ariablei + εi , where T reatmentGroupi equals one if person i was assigned into the
treatment group and zero if he was assigned into the control group, the V ariablei is listed in the
first column. A unit of observation is a survey respondent. In each regression, the sample includes
only those respondents who have finished the survey. Number of observations is 974.
divided by the number of editors of the pages of men and women are drawn from the same
distribution
One-sided hypothesis test
0.1339
0.082
0.1339
0.147
Note: The sample is restricted to the control group of the survey respondents. The null hypothesis
is that the samples are drawn from the same distribution. In row 1, the alternatie hypothesis is
that the c.d.f. of the pages of women is larger than the c.d.f. of the pages of men. In row 2, the
alternative hypothesis is that the samples are drawn from two different distributions.
43
divided by the number of editors of the pages in control and treatment groups are drawn
from the same distribution
One-sided hypothesis test: control group smaller
0.0280
0.724
One-sided hypothesis test: treatment group smaller
-0.0478
0.393
0.0478
0.710
Note: The sample is includes both treatment and control group of the survey participants. The
null hypothesis is that the samples are drawn from the same distribution. In row 1, the alternatie
hypothesis is that the c.d.f. of the pageviews of the pages chosen in control group are smaller than
the c.d.f. of the pages in treatment group. In row 3, the alternative hypothesis is that the samples
are drawn from two different distributions.
44
Appendix: Additional figures
Percentage of biographies of women
0
10
20
30
D
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013
Year
Male editors; biographies: Professionals
Female editors; biographies: Professionals
Male editors; biographies: Culture
Female editors; biographies: Culture
Figure D.1: Average (over Wikipedia editors) percentage of contributions to the biographies
of women vs men, by gender of editor and profession of the person in the biography, in 2002 –
2013
Note: The percentage is calculated as the average (over editors) percentage of the total length of
text added to the biographies of women in the total length added to all biographies. The sample of
Male editors; biographies: Professionals is restricted to male editors and a subgroup of biographies
that includes lawyers, politicians, scientists, etc; the sample of biographies: Culture includes writers,
painters, composers, etc; the exact definition of the subgroups of biographies is in Table C.1.
45
1
.8
.6
.4
.2
0
0
20000
40000
60000
80000
100000
c.d.f.: biographies of men
c.d.f.: biographies of women
Figure D.2: Cumulative distribution functions of the number of page views divided by the
number of editors from the survey editing the article, by gender (of the person in the article)
Note: Page views data is from September 2014 for the first week after the survey was conducted.
The number of page views divided by the number of editors is constructed based on the editors and
articles from the control group in the survey.
1
.8
.6
.4
.2
0
0
20000
40000
60000
80000
100000
c.d.f. of control group
c.d.f. of treatment group
Figure D.3: Cumulative distribution functions of the number of page views divided by the
number of editors from the survey editing the article, by control and treatment group
Note: Page views data is from September 2014 for the first week after the survey was conducted.
The number of page views divided by the number of editors is constructed based on the editors and
articles from the survey.
46

Gender Inequality in New Media: Evidence from

Transcription

Similar documents

Document 6482899

Veni, vidi, wikipedia - The Rowland Institute at Harvard

by Eliza Kuźnik

How to Create a Wikipedia Article

Don`t Let Them Pick YOUR Professional Pocket

Neo-Luddism and the Demonisation of Technology: Cultural

Final Project Earth Science

Lessons of the Kobayashi Maru: Cheating is Fundamental

Use of Internet in Public Library Services